📄️ Convolutional neural networks
Hands-on premium course on convolutional networks: convolution, pooling, LeNet to ResNet and MobileNet, transfer learning, Grad-CAM. 7h, applied projects, 40-question exam, verifiable certificate.
📄️ 1. Convolution and filters
Module 1 of the CNN premium course: 2D convolution computed by hand on a 5x5 image, Sobel filters, weight sharing, and why one convolutional layer has far fewer parameters than a dense layer with the same input.
📄️ 2. Stride, padding, receptive field
Module 2 of the CNN premium course: the output size formula, same versus valid padding, how the receptive field grows with depth, and dilated convolutions to enlarge it without extra parameters.
📄️ 3. Pooling and downsampling
Module 3 of the CNN premium course: max versus average pooling, GlobalAveragePooling as a replacement for the classifier head, strided convolution versus pooling, and the myth of full translation invariance.
📄️ 4. LeNet and AlexNet
Module 4 of the CNN premium course: LeNet-5 reproduced layer by layer, what AlexNet added in 2012 to win ImageNet — ReLU, dropout, GPUs, augmentation — and how to read a modern architecture table.
📄️ 5. VGG and small filters
Module 5 of the CNN premium course: why two stacked 3x3 convolutions are worth a single 5x5, VGG's repeated block, the memory cost of feature maps, and why VGG remains a reference baseline.
📄️ 6. ResNet and residuals
Module 6 of the CNN premium course: the degradation problem at depth, the residual block explained and implemented, bottleneck design, ResNet-18/50/152 and the pre-activation variant.
📄️ 7. Inception and MobileNet
Module 7 of the CNN premium course: Inception's parallel branches, depthwise separable convolutions in MobileNet, why FLOPs do not equal latency, and a first look at EfficientNet's compound scaling.
📄️ 8. Image augmentation
Module 8 of the CNN premium course: flips, crops, colour jitter, Cutout and Mixup on the waste-sorting case study, on-the-fly augmentation in the model, and what you must never augment.
📄️ 9. Transfer and fine-tuning
Module 9 of the CNN premium course: feature extraction from a frozen backbone, staged unfreezing, differentiated learning rates, the BatchNorm-in-eval-mode trap, and when transfer fails.
📄️ 10. Grad-CAM and saliency
Module 10 of the CNN premium course: Grad-CAM computed on the fine-tuned ResNet50, saliency maps, detecting a model that looks at the background, and the limits of visual explanations.
📄️ Recap and exam
Complete recap of the CNN premium course: convolution and receptive field, pooling, the LeNet-through-ResNet architecture line, augmentation, transfer learning and Grad-CAM, then the 40-question exam.